[lttv.git] / tags / lttv-0.11.3-23102008 / doc / developer / lttng-xen.txt



LTTng for Xen

Design document

Mathieu Desnoyers, November 2006

Last modification :

Only allow a single channel. Directory consumed by a lttd-xen client through
hypercalls. This design choice is made because there are no RCU lists and there
is no link between the number of CPUs seen by Xen and number of CPUs in the
dom0.

Does better than Xentrace :
Variable size records
Easily extensible event description
// not planned : missing RCU list. Multiple active traces
Reader uses poll()-like mechanism instead of polling the buffers each 100ms.


* Dom0 lttd-xen process

Does an hypercall to map the buffers (increment refcount)
Does an hypercall to poll for data, put to sleep by the OS.
Uses write from the mapped buffers to a file.
unmap buffer, decrement refcount.


* Xen

tb_control (hypercalls seen from dom0 OS)
(hypercalls done by lttctl-xen)
- trace_create
Map a page for control information, return it for mapping by user space. (rw)
Map multiple pages for buffers. return them for mapping by user space.
(shared, read-only for consumer)
increment refcount
- trace_start
Activate tracing in control information.
- trace_stop
Stop tracing in control information.
- trace_destroy
wait for pending writers (is the hypervisor preemptible ? not full.)
  Preemption : not full -> spinlock does not disable preemption : would deadlock
  for the non irq_disable case.
  volountary : poll call -> volountary. explicit __schedule call., do_yield().
Quiescent state detection.
  for_each_domain( d )
    for_each_vcpu( d, v )
      vcpu_sleep_sync(v);
decrement refcount : destroy the channels

Events coming from xen :
- Buffer switch
  through virq sent to dom0
- wakes up poll on the buffers FD.


* Dom0 Linux OS

Receive virq upon buffer switch.
xen_wakeup_readers handler : wakes up the readers for the channel.

Hypercalls are autonomous from LTTng linux kernel tracing.

on lttngxentrace create, hypercall channel_create
on lttngxentrace start, hypercall channel_start
on lttngxentrace stop, hypercall channel_stop
on lttngxentrace destroy, hypercall channel_destroy


* Facilities

Export a facilities per cpu channel too.


Q1 : In Xen preemptible ? Full ? Volountary ? Volountary X
Q2 : How to synchronize to wait for writers to finish writing ? X


Number of CPUs in dom0 != number of CPUs in Xen.
CPU hotplug in kernel tracing ? Done.
Commit	Line	Data
13ab9fcb	1
	2
	3	LTTng for Xen
	4
	5	Design document
	6
	7	Mathieu Desnoyers, November 2006
	8
	9	Last modification :
	10
	11	Only allow a single channel. Directory consumed by a lttd-xen client through
	12	hypercalls. This design choice is made because there are no RCU lists and there
	13	is no link between the number of CPUs seen by Xen and number of CPUs in the
	14	dom0.
	15
	16	Does better than Xentrace :
	17	Variable size records
	18	Easily extensible event description
	19	// not planned : missing RCU list. Multiple active traces
	20	Reader uses poll()-like mechanism instead of polling the buffers each 100ms.
	21
	22
	23	* Dom0 lttd-xen process
	24
	25	Does an hypercall to map the buffers (increment refcount)
	26	Does an hypercall to poll for data, put to sleep by the OS.
	27	Uses write from the mapped buffers to a file.
	28	unmap buffer, decrement refcount.
	29
	30
	31	* Xen
	32
	33	tb_control (hypercalls seen from dom0 OS)
	34	(hypercalls done by lttctl-xen)
	35	- trace_create
	36	Map a page for control information, return it for mapping by user space. (rw)
	37	Map multiple pages for buffers. return them for mapping by user space.
	38	(shared, read-only for consumer)
	39	increment refcount
	40	- trace_start
	41	Activate tracing in control information.
	42	- trace_stop
	43	Stop tracing in control information.
	44	- trace_destroy
	45	wait for pending writers (is the hypervisor preemptible ? not full.)
	46	Preemption : not full -> spinlock does not disable preemption : would deadlock
	47	for the non irq_disable case.
	48	volountary : poll call -> volountary. explicit __schedule call., do_yield().
	49	Quiescent state detection.
	50	for_each_domain( d )
	51	for_each_vcpu( d, v )
	52	vcpu_sleep_sync(v);
	53	decrement refcount : destroy the channels
	54
	55	Events coming from xen :
	56	- Buffer switch
	57	through virq sent to dom0
	58	- wakes up poll on the buffers FD.
	59
	60
	61	* Dom0 Linux OS
	62
	63	Receive virq upon buffer switch.
	64	xen_wakeup_readers handler : wakes up the readers for the channel.
65
66	Hypercalls are autonomous from LTTng linux kernel tracing.
67
68	on lttngxentrace create, hypercall channel_create
69	on lttngxentrace start, hypercall channel_start
70	on lttngxentrace stop, hypercall channel_stop
71	on lttngxentrace destroy, hypercall channel_destroy
72
73
74	* Facilities
75
76	Export a facilities per cpu channel too.
77
78
79	Q1 : In Xen preemptible ? Full ? Volountary ? Volountary X
80	Q2 : How to synchronize to wait for writers to finish writing ? X
81
82
83	Number of CPUs in dom0 != number of CPUs in Xen.
84	CPU hotplug in kernel tracing ? Done.
85
86